Unsupervised Clustering Methods for Lung Perfusion Data Segmentation in Electrical Impedance Tomography

In this work, we evaluated unsupervised clustering methods in segmenting the electrical impedance tomography image during the assessment of pulmonary perfusion by injection of hypertonic saline solution. In clustering the image pixels, we assume the existence of purely lung pixels (solely due to lung perfusion without effects from other organs) and hybrid pixels (which contain heart and lung effects together). We used data from 5 pigs to generate truth masks and assess the quality of clustering. Among the methods tested, the k-means with the cosine metric proved to be the best, as it obtained the 95% sensitivity median and the 90% specificity median. We prioritized minimizing the false negative cases and false positive cases, as it would overestimate regional pulmonary perfusion.